Principled Agent Debate: Adversarial Arbitration for Sycophancy Reduction in Large Language Models

A new multi-agent architecture called Principled Agent Debate (PAD) aims to reduce sycophancy in large language models, a bias where models trained with reinforcement learning from human feedback (RLHF) systematically favor agreement over accuracy [1]. The PAD framework arbitrates between two models tuned to opposing philosophical dispositions, with a pragmatist synthesizer evaluating both arguments without knowledge of their origin [1]. The architecture relies on static dispositional tuning, identity stripping before synthesis, single-round independent argumentation, and blind arbitration [1]. Researchers evaluated five prompt-based instantiations — AnCifer, DeWin, FeynStein, BurGal, and Trident — on 200 stratified questions from the SycophancyEval benchmark [1]. All PAD variants significantly outperformed a single-model baseline, which achieved just 18.5% accuracy, and an instructed-opposition baseline at 29.0% [1]. The DeWin variant reached 48.5% accuracy, a result the authors report as statistically significant with z=6.36 and p<0.001 versus both baselines [1]. The BurGal variant scored 53.0%, though the researchers describe it as an architectural validity check because its consensus/heterodox axis structurally favors the heterodox model on every benchmark question [1]. The variants did not show significant differences from each other at the sample size of 200 questions [1]. An estimated 40% of questions are affected by a pre-training floor, and the authors identify fine-tuned disposition models as the next step for development [1]. The study contributes to a growing body of work on multi-agent debate frameworks for improving model alignment, as seen in recent preprints exploring debate protocols and adversarial collaboration for language models [2][3][4].